Pablo Lanillos, CSIC robotics expert: “Metacognition is very far from AI”

Artificial intelligence is advancing at a pace that even surprises those working on its development. But, while machines gain the ability to create, learn and make decisions, there is still a boundary that separates them from human beings: understanding the physical world and reflecting on their own processes.

Pablo Lanillos, head of the Neuro AI and Robotics group and senior scientist at the Cajal Neuroscience Center of the CSIC (Superior Council for Scientific Research), investigates precisely that territory in which the human brain inspires new technologies and robots help, in turn, to understand how we think. The current limits of AI, metacognition, and the ethical challenges of a revolution require maintaining human responsibility and control to cautiously assimilate the risks of these already inevitable advances.

To what extent can a machine reproduce our ability to improvise and create?

Machines are very good in a simulated world, of words, of images, but in the physical world they are still behind humans in tasks that are not repetitive and precise. That is, they can build a car much faster than a person, but they still can’t change a light bulb properly in your house or interact with you. That’s where the real barrier is. We do experiments because we want to know how human beings have learned to relate to the physical environment, because we already see that artificial intelligence is beginning to generalize how to move to the next step, which would be invention. But at no point have we yet seen that invention as we humans do, who have the capacity as a society to invent everyone at the same time. Seeing my children makes me realize the substantial difference that still exists between machines and people.

Can a robot think about its own thinking?

It seems that metacognition, thinking about thinking, is very far from artificial intelligence. All current frontier AI companies, which are these conversational agents that we use, are looking for a way for these agents to be able to reflect on their own decisions. Because right now they are not capable. They have one entrance and one exit. On the other hand, people have the ability to know if we are sure about something because you can do something and say I’m not so sure how to do it. That uncertainty is human and machines do not know how to say I don’t know. In the past, AI did have techniques, but current models do not know how to tell you, I’m not sure, about thinking about their own cognitive process. That part of reflection is what machines still lack. And I give you an example that is wonderful: why do we count with our fingers? Children count on their fingers to help them have less uncertainty about the result.

Can AI rescue or enhance existing projects that were already considered obsolete?

Current AI is led by people who really study the brain, like Geoffrey Hinton, who is the father of ‘deep learning’ and who was awarded the Turing Award for his studies in cognitive psychology, or Demis Hassabis, from Google DeepMind and a neuroscience scholar, because we are on a path that forks and in the end it is a question of algorithm engineering, but all origins almost always begin in the brain. The study of robotics in the end is to study the human being, because delegating so much to AIs has a risk and that must always be viewed with caution because these artificial intelligences can be used for many things and we must see what their destiny is. Using them to launch rockets is not the same as using them to help you write a news story. Human responsibility must always be behind technological development because we are in a situation currently in the world, where we see how defense spending is increasing and we also see what robotics is like culturally, because there is certainly a cultural bias. In the United States, for example, robotics is more for military use. However, here in Europe it is used more in the construction and manufacturing of cars and in Japan it is understood more as a robot that is your friend. There is a very strong cultural bias, but currently there is a risk.